Topics / Benchmarks
Benchmarks
Auto topic: Benchmarks
Issues
- #6541 [Engine] strata-hy4 0.1.64: EXCEPTION_ILLEGAL_INSTRUCTION (AVX-512 EVEX) on Arrow Lake-HX — current workaround disables the 1.5x AVX-VNNI cold-expert kernels
- #6253 [Engine] macOS / Metal: open contributions, rebased on v0.2.16
- #6542 [Engine] Findings on Apple M5 Max 128 GB: Gyro-L and Hy4-Flash for coding agents on v0.2.18 (+ PRs #6538, #6539, #6540)
- #5949 [Engine] HIP/Linux: make HSA_USERPTR_FOR_PAGED_MEM=0 the default on discrete AMD cards
- #6530 [Engine] Hy4-Flash SQ-64 (strata-hy4 0.1.64): long reasoning falls into an endless repetition loop (2 of 2 attempts)
- #6531 [Benchmark] gfx1100 (RX 7900 XTX, IQ3_S): existing switches give 1.9–3.7x prompt and 1.2x decode speedup
- #6521 [Benchmark] R9700 + Threadripper PRO 3975WX: Hy4-Flash at ~30 tok/s (SQ-64, 4K) vs. 62 tok/s in the 0.2.18 notes
- #6523 [Engine] Windows NUMA in Strata: Why Our 40-Core Server Initially Used Only 20 CPU Workers
- #6526 [Engine] gfx1200 (2x RX 9060 XT, ROCm 10.1): peer-GPU fused WMMA prefill, RDNA4 MMVQ load_int_b2/Q6KTraits, and hipBLASLt fallback notes
- #6525 [Engine] Drop in MTP head replacement trained against quantized model output: +3.7 / +4.4 % decode at --spec 3
- #6339 [Engine] 0.1.62: --lookup-max-t past 8 with --draft dflash2 ends the request (dflash2: draft rows out of range); a small fix, +11..13 % on file rewrites
- #5614 [Engine] RX 9070 XT gfx1201: 0.1.58 verify unspecified launch failure triggers MES REMOVE_QUEUE GPU reset
- #6435 [Benchmark] RX 9070 (gfx1201, Windows 11, 32 GB RAM) / Gyro-S TQ1_0 / 0.2.6 → 0.2.17: 0.2.17 defaults lose ~30% prompt and ~12% decode on a 32 GB PC; two settings bring it back
- #6511 [Benchmark] 2x Tesla P40 (sm_61) split: 0.2.17 gives +9% decode, but 220K prefill is still ~5% under the 0.2.13.1 MMQ path (same-day A/B)
- #6431 [Engine] 0.1.63 STRATA_L2_PF=1 default regresses NVFP4 batch decode by 8-12% on 2x RTX 5090 with PCIe-streamed experts
- #6513 [Engine] Linux HIP: engine crashes at start (SIGSEGV in libamdhip64) with CachyOS's hip-runtime-amd 10.1 v3/znver4 builds; Arch's build works
- #6273 [Engine] 16 GB cards: guard the ~1 GB VRAM floor during KV growth and document the per-format context limit
- #6232 [Engine] 384K compaction re-reads 264K tokens (456 s): content-addressed, position-independent segment cache for near-zero-reread compaction
- #6509 [Engine] Engine 0.1.55+ decodes 35–55% slower than 0.1.52 on RTX 3080 Ti + RTX 4070 (Windows): decode expert-cache hit falls from ~92% to 55–70% although the cache gets ~40% more slots
- #6507 [Engine] q36: one early byte change costs the whole prompt — all-or-nothing match_len has no GDN state to step back to (0 reused + 38026 read, 108 s)
- #6074 [Engine] 0.1.60 prompt kernels: -13% at 20,000 tokens on a 16 GB card, +8% at 40,000, flat at 16,000
- #6483 [Engine] Long Gyro-S decode triggers HIP 719 / LiveKernelEvent 141 on RX 9070 XT
- #6503 [Engine] Swift IQ3_S_G8 (GDN Q8_0) returns garbled token soup with engine 0.1.63 on sm_120, while the same files decode correctly in another engine
- #6327 [Engine] Compaction re-reads 331K-342K tokens five times (19.5-20.1 min each): the cache budget is the blocker (PR #4338)
- #6496 [Other] Windows HIP rdna2 package (0.1.63) works on a real RX 6900 XT (gfx1030)
- #6495 [Engine] Layer-major streaming prefill on Metal
- #6487 [Engine] 0.1.62: --vision cpu fails deterministically at startup with "kv_stream: reset: an illegal memory access" (sm_86, Swift IQ3_S) — --vision gpu on the same box is fine
- #6409 [Engine] Qwen3.8-Next-Flash on Strix Halo AI Max 395 (128GB UMA): tuning for 2 concurrent streams at >100 tok/s each @ 200k context
- #6478 [Engine] Expert-cache hit rate and decode throughput regress from engine 0.1.52 to 0.1.62/0.1.63 on two 11 GB Turing cards (Qwen3.8-Flash-Next GSQ-RCO IQ3_XXS)
- #6472 [Benchmark] RTX 4090 24 GB VRAM headroom for long-context KV and IQ4 experts
- #6404 [Engine] gfx1030 (RX 6800 XT, Windows): 0.1.63 reads prompts 2x faster than 0.1.60 but decodes 11 % slower — same config, A/B on one host
- #6469 [GUI] The desktop agent's compaction request re-reads the conversation from the system prompt on Gyro (hybrid); --prompt-cache-spread cuts it from 14.9 to 2.7 s here
- #5672 [Engine] gfx1100 (RX 7900 XTX) --adapt-async 1 results in timeout followed by a display-driver reset during long generations
- #5943 [Engine] --prefill auto keeps a 5250-slot lend region on a fully-resident 3x 32 GB split (0.1.58 kept 0): decode -26%, prefill -46%; STRATA_PREFILL_LEND_PCT=0 restores the full tier
- #6466 [Engine] 0.2.14 was faster than 0.2.17 rx 9060 xt (engine 0.1.61 vs 0.1.63)
- #6429 [Benchmark] hipBLASLt tuning table by default on gfx1100 (RX 7900 XTX, IQ3_S): speed, decode and logits
- #6463 [Engine] Claude Code's small side requests evict the agent's parked conversation (default conversation cache)
- #6462 [Benchmark] RTX 4090 (sm_89): sm_86 bit-exact set +6%, sm_120 switches +4% on UD-Q4_K_XL, CPU share 32/6144 cuts 3K-token follow-up turns 26%
- #6460 [Engine] Claude Code auto-compaction returns an empty summary near the context limit unless `anthropic_thinking` is `on_request`
- #6459 [Engine] Codex CLI auto-compaction fails on /v1/responses: `compaction_trigger` input items are rejected (400)
- #6448 [Engine] Scope: plan an integrated Intel GPU (Arc B390, Panther Lake) as one pool of memory with the system in the SYCL engine
- #5678 [Engine] Qwen3.8-Flash-Next IQ3_S on sm_121 (unified memory): prefill stalls at 33,372 of 33,379 tokens, just past the 32,768 chunk; STOP not honoured, restart never finishes
- #6450 [Engine] reduce host RAM peaks when alternating parked conversations
- #5990 [Engine] V100 (sm_70) on Linux: 13× Xid 79 + PCIe CmpltTO across engines 0.1.35–0.1.59.1 — not version-specific
- #6208 [Engine] 0.2.15 decodes 28-36% slower than 0.2.7 on RTX 3060 12 GB: --expert-cache auto starts with a much worse resident set (same pack and profile)
- #5995 [Engine] Exit code 0xC0000409: stack buffer overrun on V100 + RTX A4500
- #6118 [Engine] Cold 50K prefill regression: 0.1.52 → 0.1.59 → 0.1.60 on RTX 3060 12GB (−16.5%)
- #6437 [Engine] Replies end mid-word with finish_reason "stop" at ~140K context (int8 KV, MTP on)
- #6436 [Engine] Park the engine: give all VRAM back between requests, keep the process, the context and the conversation
- #6423 [Engine] 0.2.17 sm_120 defaults regress prefill and cache-reuse TTFT on a mixed-arch layer split with the 12.0 card first
- #6421 [Benchmark] Ryzen 7 5700X, 64GB DDR4-2400, 2x RX 7800 XT 16GB (Ubuntu/ROCm) -- Qwen3.8-Flash-Next UD-IQ4_XS (256K, pinned 41.4 GiB, --spec 4)
- #6252 [Other] Intel Arc (xe driver): the installer fallback always uses --stream-experts, even with enough host RAM for a resident arena
- #6417 [Engine] 0.1.63: after one picture through the elastic image encoder, the expert cache can grow back a segment short and --prefill auto:32768 sticks at 16384 (60K cold prompt -37 %)
- #6175 [Engine] the 4-bit KV cache (q4u32) as the default on RDNA4 — scope agreement
- #6182 [Engine] Metal: --prefill auto sends 1024-token calls, so the hostpipe never overlaps the GPU (~10-12% slower prefill)
- #6413 [Benchmark] v0.2.17 on 2x Arc Pro B60 (layer split) and 1x Arc Pro B70 / Qwen3.8-Flash-Next UD-Q4_K_XL / 262K: same speed as v0.2.16, `--pipeline-windows 2` no longer produces `!!!!` (12 of 12)
- #6411 [Engine] Scope Metal host-row prefetch across calls in one prompt
- #6412 [Engine] Scope opt-in IQ2_S and IQ3_S Metal prompt kernels for M5 Ultra
- #6410 [Engine] Scope an opt-in compact two-slot mode for resident Metal
- #6392 [Engine] Linux HIP RX 7900 XTX (32 GB RAM, IQ2_XS low-RAM): engine hangs within 4-10 min of sustained serving, verify timeouts, also on 0.1.63
- #6088 [Engine] Nvidia v100 performance slow down -12% with "own kernels"
- #6244 [GUI] Changing an option on the model screen silently drops drafting: dflash2 and resets STRATA_RESIDENT_PIN to 0
- #6312 [Benchmark] `strata bench run` marks a v0.2.16 git checkout nonconforming: the commit lookup checks `terminal/.git`
- #5551 [Engine] q36 on the qwen35moe 19B-A3B shape: three localized walls with clean data
- #6156 [Engine] Qwen3.6-35B-A3B on 2x RTX 3060: qwen35moe ignores multi-GPU flags — patch adds a card-1 expert tier, 37.6 → 74.8 tok/s (+97%)
- #6052 [Engine] Equal prompt chunks skip periodic conversation checkpoints (--prompt-cache-every)
- #6403 [Engine] 12 GB of VRAM on an RTX 5080: --prefill auto picks a 7,936-token chunk where 8,192 fits; 16K prompts read in three pieces, 16-30 % slower
- #6269 [Engine] gfx103x: prompt throughput gaps left by the MMQ removal (dense projections, IQ4_XS down)
- #6265 [Engine] layer-split auto: the #2644 clamp picks a slower split on asymmetric same-VRAM pairs (measured +10% hand-pinned)
- #6264 [Engine] RDNA2 gfx1030: 37% prefill regression persists in 0.1.62 — MoE GEMM gate/up 3.4x slower than 0.1.59
- #6261 [Engine] Run one Flash-Next model across two DGX Sparks
- #6174 [Docs] Mac: source setup gives Q2_0 the DFlash2 drafter, but MAC.md says only the two Unsloth sets get it
- #6173 [Other] Source setup accepts --context 262144 for Qwen3.6-35B-A3B on a 12 GB card, and the engine then refuses to start
- #6136 [Engine] 0.1.60: prompt + appended tail re-reads a whole equal chunk (append 6.4x slower than 0.1.54, first read 7.5% faster)
- #6130 [Engine] RX 7900 XTX (gfx1100, Linux), 0.2.15: picture text misread ("STRATA 412" for "STRATA 42") after earlier requests on the default HIP build; the -DSTRATA_MMQ=ON build reads it right
Pull requests
- #4310 [Engine] session files: format v2 with per-stage save/restore under --layer-split
- #6520 [Engine] Inkling-2-Mini from the official NVFP4/MXFP8 checkpoint, expert offload and CPU experts for 8-16 GB cards
- #6023 [Engine] peer_stream_ring: fix >= 2K peer_i8 ring wraparound + V100 docs
- #6210 [Engine] Opt-in P2P q8_1 decode push, fused P2P down-projection, and GPU peer sync flag
- #6529 [Engine] Volta sm_70 decode optimizations, zero-VRAM in-place Q8_0, NVLink split head, and volta-forever detection
- #6548 [Engine] Add opt-in streamed-KV QSA slot-group batching, pre75 multi-slot attention, and batch SIDE_MIXER
- #6547 [Engine] Add opt-in multi-slot batched MTP drafting (STRATA_BATCH_MTP_FUSED)
- #6546 [Engine] Opt-in resize_for_overwrite and parallel host memcpy for kv_mode != 0 park/restore
- #6545 [Engine] Size STRATA_PARK_ASYNC VRAM stage by device-resident bytes under kv_mode != 0
- #6544 [Engine] Fix O(N^2) scan in getline_fd and avoid full-line copy in BGEN/parse_i64_list
- #6540 [Engine] Hy4-Flash: SQ-96 starts on 128 GB Macs, --info no longer crashes, session reuse on by default, reasoning_effort max in configs
- #6539 [Engine] Hy4-Flash (Metal): resume at the longest common prefix, several conversation slots, SSD tier kept across restarts
- #4306 [Other] serve: session ladder — conversations survive an engine unload
- #3773 [Engine] q36: MTP drafts for sampled requests behind STRATA_Q36_MTP_SAMPLED (stacked on #3678)
- #6535 [Benchmark] RTX 5060 Ti / Qwen3.8-Flash-Next / IQ3_S / strata-qwen38-thinking-v1-draft.1
- #6532 [Engine] prefill: opt-in STRATA_I8_DEFER - unpicked streamed entries no longer cut the int8 expert groups into fragments (gfx1030 4K prompt +3.6%)
- #3800 [Engine] Token mask for JSON response formats: MTP drafts survive constrained decoding
- #5436 [Engine] opt-in STRATA_ARENA_NUMA: place the expert arena per NUMA node on two-socket Linux
- #6506 [Engine] add Inkling 2 model family and dual-GPU looped MoE engine support
- #6399 [Engine] Prefetch routed expert weight rows into L1 on GB10
- #3607 [Engine] native Windows support for Intel Arc SYCL, rebased onto 0.2.18 (engine 0.1.64)
- #6365 [Engine] Prefer FP32 TLA sparse prefill on Battlemage in builds with local TLA headers
- #6514 [Benchmark] RX 7600 XT (gfx1102): ROCm 7.2.4 vs 10.1, and the CachyOS ROCm 10.1 crash
- #4928 [Engine] long --lookup-max-t windows over a layer split (opt-in, single conversation)
- #6510 [Engine] --experts-wex4: read the routed experts from a checked WEX4 container (opt-in, file tier)
- #4619 [Engine] DS4_LOOKUP: draft a round from the request's own history (opt-in)
- #6197 [Engine] int8 expert kernels: IQ4_XS as a down format
- #6468 [Benchmark] RX 9070 XT on Windows, Gyro-S at 131K: app 0.2.17 / engine 0.1.63 with a tuned config (standard 12/12)
- #6176 [Engine] q4u32: the 4-bit KV cache with a fitted step as the default on RDNA4
- #6498 [Benchmark] RTX 3060 12GB / Qwen3.8 Flash-Next Q2_0 / standard v1 draft.1 + 260K xhigh
- #4255 [Other] serve: refuse a port another program holds on 127.0.0.1, and say which program
- #6419 [Engine] autonomous hardware-synchronized dual-GPU decode window (P2P zero-wait) [Draft speed Lab 230 tok/s+ RFC]
- #6477 [Benchmark] Arc Pro B60 / Qwen3.8-Flash-Next / IQ3_S / strata-qwen38-thinking-v1-draft.1
- #5476 [Engine] gfx103x decode: opt-in one-warp MMVQ for K <= 4096 and a per-format expert layout table (STRATA_MMVQ_NW1, STRATA_EXP_TABLE)
- #5475 [Engine] opt-in int8 copies for the persistent hc read (STRATA_HC_PERSIST_Q8, STRATA_HC_I8ONLY)
- #2843 [Engine] Slot checkpoints for clients that rewrite the end of their history (3/3, depends on #2527 and #2528): STRATA_SLOT_CHECKPOINTS, STRATA_BRANCH_CHECKPOINT, BCHECKS
- #2528 [Engine] Slot placement policy for STRATA_SLOT_CACHE_AFFINITY (2/3, depends on #2527): protection by history length, retention scores, value decay, no wait
- #2527 [Engine] Keep conversations in pipelined batch slots (1/3): active-row windows, park=, cache-aware placement behind a policy interface
- #1614 [Engine] shared KV pool for all lanes (--kv-pool-tokens): at 2 x 256k the first token comes 2-5x sooner and lane moves ~10x faster; off = identical to main
- #6464 [Benchmark] Strix Halo 64 GB unified memory (Ryzen AI Max+ 395)
- #6453 [Engine] SYCL: an integrated Intel GPU is planned as one pool of memory with the system (Arc B390)
- #6386 [GUI] isolate launch configs for concurrent tuning commands
- #6387 [TUI] recheck saved tuning when its launch conditions change
- #6458 [Engine] STRATA_FILL_STAGED: copy mapped blobs to the GPU via pinned buffers
- #6455 [Engine] SYCL: an Xe3 GPU (Panther Lake, Arc B390) runs the Battlemage kernels; STRATA_INTEL_GEN sets the class
- #3919 [Engine] kv: size the resident ring from the run's depth (opt-in, STRATA_KV_RING_ADAPT)
- #3729 [Engine] SYCL: flag Intel iGPUs as integrated so the expert cache is sized from MemAvailable
- #3648 [Engine] --peer-device host route: STRATA_PF_PEER_PIECES=1 (opt-in) moves the activations and the sums in 8 pieces (+7 % prompt on 2x V100)
- #6443 [Engine] Avoid redundant top-K scans in Metal expert read-ahead
- #6434 [Engine] Use eight-row tails in M1 resident expert GEMMs
- #6432 [Engine] Reuse Metal resident prefill scratch buffers
- #2427 [Engine] HC_Q8: skip the layers' bf16 hyper-connection copies on a layer split too
- #4766 [Engine] gfx906: build dflash2 for Radeon VII
- #6430 [Engine] q36: --kv fp32|fp16|fp8|fp4 for the K/V cache — 262,144 tokens on one 12 GiB card
- #6425 [Engine] q36: card-1 expert tier (--peer-device) for Qwen3.6-35B-A3B, 37.9 → 74.8 tok/s
- #4019 [Engine] Mac: calibrate the decode kernels' tiles on the device (setup.sh --calibrate)
- #6433 [Engine] Fix: read the MTP draft layer's expert count from experts.bin, to support models with other expert counts such as pruned models
- #2784 [Engine] HIP: the RX 6600 / 6650 XT (gfx1032) in setup's cards
- #6263 [Engine] Add opt-in MPI/NCCL split inference for two DGX Sparks
- #3819 [Engine] Layer split: a stage's commit without the host wait, as on one GPU (`STRATA_SPLIT_COMMIT_ASYNC=1`, opt-in)
- #3944 [Engine] --pipeline-windows: the front stages' GDN snapshot by buffer rotation instead of a copy per window, same bits (opt-in, `STRATA_PIPELINE_SNAP_RING=1`)
- #3760 [Engine] A third decode window in flight on a layer split (`--pipeline-windows 3`, opt-in): a host-side scheduler, a slot pool, per-stage undo, an 11-output draft chain
- #3678 [Engine] q36: quantized KV caches behind --kv (int8 / fp8 / rk4v4-e8 q4)
- #6418 [Engine] Leave STRATA_BATCH_SELF_COMMIT out of the Linux CUDA parallel defaults when STRATA_PIPED_MTP is asked (#6160)
- #6183 [Engine] Metal: --prefill auto hands each call 4 prompt chunks so the hostpipe overlaps the GPU
- #6414 [Engine] gfx906: reference build for MI50/MI60/Radeon VII (0.2.11 + int6 kernels)
- #6415 [Engine] Skip older draft rows in resident Metal prefill
- #5562 [Engine] prefill: the fused layout's shrink asks every layer's path, not only the formats
- #4883 [Other] Setup: --gpus on Windows AMD: RDNA 2, 3 and 4 cards sharing one model
- #6314 [Benchmark] RX 6700 XT and RX 6600 (x1, x2, with and without a rocBLAS table) / Qwen3.8-Flash-Next / IQ3_S / strata-qwen38-thinking-v1-draft.1 on the 0.2.16 rx6000 package
- #4859 [Engine] a 3-bit Lloyd-Max K/V (--kv oct3) with an exact fp16 window, and the gfx1100 decode ports
- #6408 [Docs] Community benchmark: one Radeon VII — ROCm 10.1 vs 7.14, engine 0.1.63 vs 0.1.62
- #6406 [Engine] dflash2: unlock spec 8 window for desktop Blackwell cc 12.0 cards >= 15 GB (+89% decode on code review)
- #5595 [Engine] q36: instantiate the grouped down-expert kernel for Q4_K (fmt 12)
- #4272 [Engine] Metal: opt-in routed-expert tiles (STRATA_METAL_MOE_GU/_DN); Q2_0 8,1,2 + 4,2,4 is 17-24% faster on M1 Max